Top AI Model Tracker

Which lab is shipping the best frontier model? Polymarket's leaderboard books settled at the end of June and it has not relisted, so the live question trades on Kalshi at three horizons at once—and all three moved this week.

The year-end book reversed. KXLLM1-26DEC31 now prices Claude at 68.15%, up from 59% three days ago, with Grok collapsed from 23.5% to 11.15% and ChatGPT at 11.2%. The challenger trade that dominated this page has been unwound: Grok's leg took $7,664 today against $40,406 on Friday, and it now sits behind ChatGPT rather than ahead of it.

And the money went somewhere stranger. The deepest leg on a $51,959 book is not the leader and not the challenger. It is Ernie at 0.25%, taking $14,230—twenty-seven percent of the day's flow on a Baidu model the same book prices at four-hundred-to-one. Kimi at 0.5% took $3,141 and Qwen at 1.2% took $2,939. Roughly $20,000, forty percent of the book, changed hands on legs priced at or under 1.2%.

The weekly board is still a single-vendor sweep, and still not where the money is. KXTOPMODEL-26AUG24 prices claude-opus-5-max at 84.5% and claude-opus-5-high at 15%. The only non-Anthropic name with a live leg is qwen3.8-max at 0.5%—and it took $3,538 of the $5,230 book, sixty-eight percent of the week's flow.

Two horizons, same behavior: price the incumbent, buy the long shot.

The December book alongside the week ending August 24. Right now they name the same leader at very different confidence: 84.5% this week, 68.15% by December—a gap that has narrowed sharply from 90% versus 59% on Friday.

That narrowing came from the far end. The year-end book did not become more confident because the weekly board changed; it became more confident because Grok halved. Open interest tells you the position was inherited rather than fresh—Claude's leg carries 1,004,645 contracts, ChatGPT's 960,578, Grok's 836,396—so a twelve-point move on $7,664 of volume is a repricing of a very large book by a very small trade.

The two legs that define the year-end book, and they are not the two you would expect. Claude at 68.15% took $11,289. Ernie at 0.25% took $14,230—more than the leader, on a leg priced at a quarter of one percent, carrying 443,020 contracts of open interest.

This is the same structure that showed up in every commodity book on the platform today: gold's deepest rung priced at 5%, copper's at 0.5%, ETH's at 6.5%, gasoline's at 1.5%. Here it is in AI. Whatever explains it is not asset-specific, and on a question about which lab ships the best model in December it is harder to call hedging than it is on an oil ladder.

The August monthly and the year-end coding book. KXLLM1-26AUG31 gives Claude 94.5% with thirteen days left—but its deepest leg is ChatGPT at 2.5% taking $2,852, sixty-nine percent of that book. Third horizon, third instance of the same pattern.

Coding is where the top slot has changed hands most often in 2026. KXCODINGMODEL-26DEC prices Anthropic at 58.5%, OpenAI at 28.5%, xAI at 14%—a genuinely three-way book, and the only contract on this page where the incumbent is under two-thirds. It is the leading indicator to watch: if a challenger is going to break through anywhere, it breaks through in coding first.

KXTOPAI-27 prices which companies hold a top-ranked model at any point during the year, so the legs do not sum to 100%. OpenAI leads at 35%, xAI 17.5%, Moonshot AI 14%, Meta 13.5%—a wide distribution on breadth next to a year-end book that gives one lab better than two-thirds. The market expects the lead to change hands repeatedly without settling.

Alongside it, the coding-specific weekly, which is worth reading for what it is rather than what it says: seven of its nine legs price at exactly 1.5% and carry exactly $200 each. That is a seeded book, not a traded one. Only Claude at 89.5% and Kimi at 4.5% have independent flow, and Kimi's $528 is the deepest leg on it.

• Grok's unwind is the story. 23.5% to 11.15% in three sessions, and the $40k that made it the deepest position on this page is gone. The challenger thesis this workspace has tracked since May did not play out—it was sold.
• $14,230 on Ernie at 0.25%. The largest single position in AI on this platform is on a Baidu model priced at four-hundred-to-one, and it is not alone: forty percent of the year-end book traded on legs under 1.2%. Watch whether this rolls or decays.
• 68% versus 84.5% is the handoff premium. The market gives Claude better than five-in-six to hold the board through Sunday and roughly two-in-three to hold it through December. That gap was thirty-one points on Friday and is sixteen now.
• Coding is the only competitive book. 58.5/28.5/14 across Anthropic, OpenAI and xAI. Every other contract here is a one-lab sweep with lottery tickets attached.
• Release windows drive everything. Anthropic, Google, OpenAI, xAI and the Chinese labs (Alibaba, Baidu, ByteDance, Moonshot, Z.ai, DeepSeek) reprice these books on flagship launches. The weekly moves within hours; the year-end book takes days—except when it moves twelve points on $7,664, as it just did.
• The weekly rolls every Monday. Books for the week ending August 17 settled and Kalshi relisted for August 24. Expect this anchor to roll on each custodian run.
• Polymarket has been out of this category since June. No relisted book, so there is no cross-exchange spread here—unusual for a topic with this much retail attention.